Balancing SFT and RLVR in training video-LLMs

Determine the appropriate balance between supervised fine-tuning and reinforcement learning with verifiable rewards when training video large language models to achieve robust spatio-temporal video understanding.

Background

The paper empirically examines training paradigms and finds complementary strengths for supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR), including task-dependent advantages.

Despite these observations, the authors highlight that the optimal balance between SFT and RLVR for video understanding remains unresolved, identifying it as an open question for future research.